Skip to content

test: six-client connection matrix (Phase 1 exit bar) #223

Closed

Description

@ifahimreza

Phase 1's exit bar: a scripted connect → read → gated write passes on a clean install in each client.

Client Auth path
ChatGPT OAuth
Claude (web / desktop) OAuth
Claude Code App Password, then OAuth after the loopback fix
Codex App Password
Gemini CLI App Password
Cursor App Password
  • A written script (the prompts plus what each step must show)
  • Results table in tests/CLIENTS.md with the date and versions
  • Failures filed as their own issues

Activity

  1. ifahimreza commented on Oct 4, 2026

    @ifahimreza
    ContributorAuthor

    #324 adds scripts/client-matrix.mjs and tests/CLIENTS.md. On Playground the key path passes for Claude Code, Codex, Gemini CLI and Cursor: initialize, initialized, tools/list, prompts/list, a read, and a gated write with --write. Two things are still to do before this exit bar is met:

    • run the sign-in path on an HTTPS site with sign-in on (--token);
    • connect the six real apps once, and fill in the results table in tests/CLIENTS.md.
  2. ifahimreza commented on Oct 5, 2026

    @ifahimreza
    ContributorAuthor

    Real-app pass started 2026-10-05 on the 1.5.1 zip from WordPress.org (fresh Playground).

  3. ifahimreza commented on Oct 13, 2026

    @ifahimreza
    ContributorAuthor

    Progress 2026-10-14 (recorded in #366):

    Left for this exit bar, each needing the owner's own sign-in:

    1. ChatGPT: Developer mode → add a connector with https://saddle.to/wp-json/saddle/v1/mcp, approve on the consent screen, then read and preview a trash. Plus/Pro accounts can only read through custom apps, so the write row may stay N/A unless the account is Business.
    2. Cursor: ! cursor-agent login, then read and the gated write.
    3. Gemini CLI: npm i -g @google/gemini-cli, sign in with Google, then the same three steps.
  4. ifahimreza commented on Oct 14, 2026

    @ifahimreza
    ContributorAuthor

    Phase 1 exit bar met, 2026-10-14. Fahim ran the remaining real-app tests himself (ChatGPT, Cursor, Gemini CLI) and reports them done. Agent-run rows are in tests/CLIENTS.md (#366): the script passes all six, and Claude Code, Codex and Claude (claude.ai connector) pass with the real app.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions